SSChNet: Self Supervised Learning based pre-trained network for Child Speech Recognition

Automatic Speech Recognition (ASR) is a key part of human-computer interaction, playing a major role in education, accessibility, and personal assistant technologies. However, recognizing child speech remains a significant challenge due to the scarcity of large labeled datasets and the distinctive traits of children’s voices, such as higher pitch, greater tonal variation, and frequent mispronunciations. ASR models predominantly trained on adult speech often struggle to process child speech accurately. To address these challenges, we propose a modified framework based on the Wav2Vec2 architecture, which leverages self-supervised learning to pre-train on a large set of unlabeled child speech data, followed by fine-tuning on a smaller labeled dataset. This approach incorporates targeted enhancements, including pitch normalization, phoneme-aware masking, and a child-specific speech dataset, to better capture the unique characteristics of child speech. Our investigation show significant performance improvements, with the large model achieving a WER of 8.13% on MyST_test and 14.87% on PFS_test, demonstrating its ability to handle diverse and challenging datasets effectively. This framework offers a promising solution for developing ASR systems tailored for children, with potential applications in education, child-computer interaction, and accessibility tools.